Predicting Audience Interests from Social Media Captions Using a Semi-Supervised Approach

The aim of this paper is to predict the interests of social media influencers’ audiences. This paper presents a novel approach to the multi-label classification of interests in six categories: sports and fitness, travel, fashion, electronics, photography, and food. Out of the five classifiers, the Multilayer Perceptron (MLP) achieves the highest accuracy of ${9 8 . 2 2 \%}$ on the single-labeled dataset. Regardless, evaluation solely based on single labels only partially captures the complexities of multilabel classification. The Random Forest Classifier appears the most accurate among glass box models in labeling sentences belonging to one category and also showed promise in categorizing captions belonging to multiple categories. On the other hand, the SVM and KNN classifiers struggle with multi-label classification and often mislabel or fail to capture the labels in complex sentences. Despite lower accuracy on the single-labeled dataset, the Gradient Boosting Classifier demonstrates the most promising performance in labeling complex sentences, whereas, although achieving the highest accuracy for single-labeling, the MLP faces challenges with complex sentences containing multiple categories, indicating the need for a significantly larger dataset to improve its performance.